Just to add to this:
You’ll find that the AFD in GM will be lower than in WM using higher b-values (nearer 3,000s/mm²), and these noisy lobes in the GM would be attenuated – another good reason to go for high b 
However, you might consider the use of a 2-tissue multi-tissue CSD approach here – we’re starting to recommend that for all single-shell analyses, definitely worth a try. General instructions can be found here – only change is that in the dwi2fod call, you should only pass the WM and CSF arguments (one for the input response, one for the output ODF image, for each tissue type), and skip the GM ones. That should already attenuate the problem to some extent (although not completely).